I am following the simple example of fitting a line.

I used slightly different sample points than in the tutorial and the training completly diverges to infinity. Here is the working code that is exactly as in the tutorial:

using Flux
using Statistics

Flux.Random.seed!(42)
f(x) = 4x + 2

m = Dense(1=>1)
loss(model, x, y) = mean(abs2.(model(x) .- y));

using Flux: train!
opt = Descent()
x = hcat(0:5 ...)
d = [(x, f.(x))]
train!(loss, m, d,opt); loss(m, x, f.(x))
m.weight, m.bias
for epoch in 1:200
    train!(loss, m, d, opt)
end
@show loss(m, x, f.(x))

however when I use a slightly different set of sample points, for example x = hcat(1:6 ...) or x = hcat(0:6 ...) or x = hcat(1:10 ...) or anything else than in the tutorial the loss from the train! function becomes larger and larger diverging to infinity.

I would expect the training to be more stable, especially considering that this is an extremly simple model.

(-1:5 converges, -1:6 diverges, 0:8 diverges, ...)